Accurate speech-to-text, at half the going rate.
Batch and streaming transcription through one OpenAI-compatible API. Proprietary and open-weight models behind a single integration, with speakers, entities and language-model processing available on the same request.
What a speech-to-text API actually has to do.
Speech-to-text — also called automatic speech recognition, or ASR — converts spoken audio into written text. A production API has to do more than that: it has to return when each word was spoken, hold up on real-world audio with overlapping speakers and background noise, stay fast enough for the product it sits inside, and cost little enough that transcribing everything is affordable rather than a budget line you ration.
Most teams discover the gap after integrating. Transcription alone is rarely the product. You need to know who spoke, pull out names and dates, judge tone, and hand structured data to something downstream. That normally means a second vendor, a second integration and a second bill.
One request, the whole pipeline
Ultrafield exposes transcription as the first stage of a composable pipeline rather than a standalone endpoint. Diarization, speaker identification, transcript formatting, entity detection, sentiment and language-model processing attach to the same job. Stages are metered independently, so you pay for what a request runs and nothing else.
Economical without being slow or inaccurate
Being cheaper usually means a smaller model and worse output. We take a different route: run a lean inference stack, price per processed second rather than per seat or bundle, and let you choose the engine that fits each workload. Use hosted Whisper large-v3-turbo where latency matters, large-v3 where accuracy matters most, and Ultrafield ASR-1 where you want our own model on difficult audio. Faster token processing and lower latency ceilings are available on Pro and Enterprise, so scaling up buys speed rather than just a bigger invoice.
From audio to structured output.
Upload a file, pass a remote URL, or open a stream. One request shape covers all three.
Name the engine explicitly. Accuracy, latency and cost are yours to trade off, not ours to guess.
Diarization, speaker ID, formatting, entities and sentiment run on the same job, not a second integration.
Text, word-level timestamps, speaker turns and derived fields come back in one normalized response.
Choose the engine. Keep the API.
Per processed minute.
Where teams put it to work.
Transcribe and diarize large call archives, then extract entities and sentiment in the same pass.
Turn recordings into formatted, speaker-attributed transcripts and structured summaries.
Stream audio and render interim results immediately, finalizing segments as they settle.
Low-latency transcription in the loop of a conversational product, with interruption handling.
Word-level timestamps for caption files, alignment and searchable archives.
Long-form interviews and regulated recordings, with retention you control.
Questions, answered.
How much does the speech-to-text API cost?
Batch transcription starts at $0.0012 per minute ($0.07 per hour) on Whisper large-v3-turbo, and $0.0021 per minute ($0.13 per hour) on Ultrafield ASR-1. Streaming starts at $0.0018 per minute. You pay for the stages a request actually runs, billed per second with no minimums.
Which speech recognition models can I use?
Ultrafield ASR-1, our proprietary model, alongside open-weight Whisper large-v3 and large-v3-turbo. You name the model in the request, so switching engines never means switching integrations.
Does it support both batch and real-time transcription?
Yes. Uploaded files, remote audio URLs and low-latency streams share one API surface. Batch suits archives and analytics; streaming suits live captions and conversational products.
Is the API compatible with OpenAI?
The audio endpoints follow OpenAI conventions for routes, Bearer authentication, errors and usage reporting, so most clients need only a base URL and key change. Ultrafield-specific capabilities are added as extensions rather than by altering the compatible surface.
What languages are supported?
Language coverage follows the model you select. The Whisper family is multilingual across a broad set of languages; consult the models documentation for the current list and per-language guidance.
Do you train on customer audio?
No. We never train on customer audio, transcripts, derived content, voiceprints, prompts or outputs, and we do not route customer data to downstream services that may train on it.
How long is audio retained?
Source audio is removed within 24 hours of terminal processing, and derived result content is retained for 30 days by default. Operational and billing records are kept separately from customer content.
Can I get lower latency and higher throughput?
Yes. Pro raises concurrency to 50 batch jobs and 25 streams with priority job scheduling. Enterprise adds custom concurrency, throughput and rate limits, dedicated hosting and capacity planning for burst volume.
Give your product better ears.
Move from raw audio to accurate transcripts, speakers, entities, sentiment and application-ready output through one pipeline.
Batch and streaming · Proprietary and open models